|
|
MIT Electrical Engineering and Computer Science
EECS Event |
Monday, April 30, 2001
1:30 PM (refreshments 1:15)
RLE Conference Room, Room 36-428
EECS Special Seminar
Abstract
Sensory information processing systems operate all around us under physical constraints. These physical constraints include fundamental and phenomenological sources of noise, limitations on available resources, and performance requirements. Any system exhibits tradeoffs between performance (speed and accuracy) and cost (power and size).
In this talk I describe an approach towards quantifying these tradeoffs for natural and engineered systems. This approach leads to an understanding of the performance limitations and suggests strategies for improving performance. In the context of sensory information processing systems, I investigate channel capacity as a measure of performance and power as a measure of cost. The ratio of capacity to power defines the bit-energy, the minimum energy required to transmit a single bit of information.
I consider several physical systems in detail: the blowfly photoreceptor, an adaptive silicon photoreceptor, and a nonlinear CMOS inverter. For each system I have developed detailed models that describe signal transfer, noise, and power dissipation. The blowfly model produces unique predictions about the biological noise sources that limit neural information transmission. I compare the performance of this biophysical photoreceptor with that of a silicon photoreceptor. My results indicate that the blowfly photoreceptor has higher capacity at low light levels, but the silicon photoreceptor can be biased to operate more efficiently. I also find that, in the absence of noise, the classical CMOS inverter can be operated with arbitrarily low bit-energy.
I'll conclude with discussion of possible applications and a description of ongoing and future work.